Most AI conversations in digital commerce right now are about adding AI on top of what already exists. A chatbot here, a recommendation widget there, a co-pilot floating in the corner waiting to be clicked. The underlying interface, however, stays the same. The menus, the filters, the search bar, the click-heavy journey, all of it remains unchanged.
What if AI didn’t sit on top of the experience, but became the experience itself? That’s the question we set out to answer along with Finnish car dealership Saka and partner Contentful.
Together, we built a working proof of that idea.
TL;DR:
AI becomes genuinely powerful when it’s a true part of the site rather than a layer on top of it. Together with Saka and Contentful, we built a live demonstration of what that looks like in practice. Watch the full keynote below.
The problem with bolting AI onto old interfaces
Websites were designed for a page-based world. You navigate, you filter, you click. That logic is so deeply baked in that most teams don’t question it, they just add AI capabilities into the existing structure and call it progress.
But there’s a ceiling to that approach. The interface is still a menu. The user is still doing the work and AI in that setup is a layer, not a leap.
Buying a car is a good stress test for this. It’s one of the biggest financial decisions most people make involving dozens of variables such as fuel type, total cost of ownership, safety specs, family fit, personal preference. Very few of those variables translate cleanly into a filter dropdown and a normal search interface handles maybe 20% of what actually matters to a buyer.
Saka knows this well. They sell around 50,000 cars a year across 32 locations in Finland, and one in three of those sales happens entirely online.
What we built
We didn’t redesign the Saka website. We didn’t touch the backend, the inventory, or the data layer. Instead what we built was an entirely different way of interacting with the site.
The agentic version of the site replaces menus and filters with conversation. You can talk to it, you can show it photos and it remembers who you are, builds a picture of your preferences over time, and it acts on your behalf. It doesn’t just suggest, it does.
Contentful provides the structured, composable content backbone that makes this possible. For an AI agent to navigate a catalog, answer questions accurately, and take actions without hallucinating specs or accessories, the content underneath needs to be clean, structured, and API-accessible. That’s the foundation everything else runs on.
When AI becomes the interface: what it looks like in practice
- Conversational browsing with no clicks. Instead of manually filtering through thousands of options, a user can simply say “show me only Toyotas.” The agent filters, compares models, explains price differences, answers questions about specific features, and navigates back and forth — all through natural language. No cursor required.
- Image-based discovery. Point a phone camera at a car interior, and the agent identifies the model, describes the upholstery, and cross-references it against the user’s known preferences. If you don’t know what a spec is called but you know what it looks like, that’s enough.
- Memory and continuity. Purchase journeys rarely happen in a single session. An agentic system remembers what a user has browsed, what they’ve favorited, and what they’ve said they don’t want. Say “I don’t like Mercedes” once, and the system excludes Mercedes for the rest of the journey.
- Actions, not just answers. This is where agentic really separates itself. A user wants to book a test drive for a specific car on a specific afternoon. The agent checks availability, confirms the time, and books it. No form to find, no calendar to navigate. A callback request for a second car gets captured in the same conversation.
Answering questions is useful. Taking responsibility for an action is a different thing entirely.
Why this is bigger than car retail
Saka is a strong use case precisely because it’s hard. Complex inventory, high-consideration decisions, long purchase journeys that span multiple sessions and channels, and real commercial consequences if the AI gets something wrong. Saka’s CIO was direct about this: if the agent tells a buyer a car has a feature it doesn’t have, that’s false advertising, and that’s a returns problem.
But the same pattern applies to any business where the catalog is complex, the decision is high-stakes, and the customer journey is long. B2B procurement, financial products, healthcare, travel — anywhere that forms and filters are a poor substitute for a real conversation.
The infrastructure requirements are real. You need structured content, API-first architecture, and a platform built to deliver across channels. You need an agent that can observe, reason, remember, and act — not just respond. And you need to be able to trust what it says.
Those aren’t small things. But none of them are theoretical anymore.
Curious what this could look like for your business?
